Visualization of SARS-CoV-2 Infection Scenes by 'Zero-Shot' Enhancements of Electron Microscopy Images

Abstract

Electron microscopy (EM) recordings of infected tissues serve to diagnose a disease, and they can contribute to our understanding of infection processes. Consequently, a large number of EM images of the interaction of SARS-CoV-2 viruses with cells have been made available by numerous labs. However, due to EM recording techniques at high resolution, images of infection scenes are very noisy and they appear two dimensional ( flat). Current research consequently aims (A) at methods that can remove noise, and (B) at techniques that allow for recovering a 3D impression of the virus or its parts. Here we discuss a novel method which can recover a spatial impression of a whole infection scene at high resolution. In contrast to previous approaches which aim at the reconstruction of single spike proteins or a single virus, the here used method can be applied to a single noisy EM image of an infection scene. As one example image, we show a high resolution image of SARS-CoV-2 viruses in Vero cell cultures (Fig. 1). The method we use is based on probabilistic machine learning algorithms which can operate in a zero-shot setting, i.e., in a setting when just one noisy image (and no large and clean image corpus) is available. The probabilistic method we apply can estimate non-noisy images by inferring first order statistics (pixel means) across image patches using a previously learned probabilistic image representation. Estimating higher order statistics and appropriately chosen probabilistic models then allow for the generation of images that enhance details and give a spatial impression of a full nanoscopic scene.